dany/libwebp
1
0
Fork 0
mirror of https://github.com/webmproject/libwebp.git synced 2025-07-13 06:24:27 +02:00
Code Issues Actions Packages Projects Releases Wiki Activity

fix some SSIM calculations

Browse Source 
* prevent 64bit overflow by controlling the 32b->64b conversions
  and preventively descaling by 8bit before the final multiply
* adjust the threshold constants C1 and C2 to de-emphasis the dark
  areas
* use a hat-like filter instead of box-filtering to avoid blockiness
  during averaging

SSIM distortion calc is actually *faster* now in SSE2, because of the
unrolling during the function rewrite.
The C-version is quite slower because still un-optimized.

Change-Id: I96e2715827f79d26faae354cc28c7406c6800c90
This commit is contained in:
Pascal Massimino
2016-10-04 07:10:27 +02:00
parent 51b71fd295
commit ba843a92e7
3 changed files with 103 additions and 77 deletions
Download Patch File Download Diff File Expand all files Collapse all files

75
src/dsp/enc.c
View File

@ -693,31 +693,35 @@ static void Copy16x8(const uint8_t* src, uint8_t* dst) {
//------------------------------------------------------------------------------
// SSIM / PSNR
static const double kMinValue = 1.e-10; // minimal threshold
// hat-shaped filter. Sum of coefficients is equal to 16.
static const uint32_t kWeight[2 * VP8_SSIM_KERNEL + 1] = {
1, 2, 3, 4, 3, 2, 1
};
static const uint32_t kWeightSum = 16 * 16; // sum{kWeight}^2
static WEBP_INLINE double SSIMCalculation(
const VP8DistoStats* const stats, uint32_t N /*num samples*/) {
const double w2 = N * N;
const double C1 = 6.5025 * w2;
const double C2 = 58.5225 * w2;
const double xmxm = stats->xm * stats->xm;
const double ymym = stats->ym * stats->ym;
const double xmym = stats->xm * stats->ym;
const double sxy = stats->xym * N - xmym;
const double fnum = (2 * xmym + C1) * (2 * sxy + C2);
double fden;
double sxx = stats->xxm * N - xmxm;
double syy = stats->yym * N - ymym;
// small errors are possible, due to rounding. Clamp to zero.
if (sxx < 0.) sxx = 0.;
if (syy < 0.) syy = 0.;
fden = (xmxm + ymym + C1) * (sxx + syy + C2);
return (fden != 0.) ? fnum / fden : kMinValue;
const uint32_t w2 = N * N;
const uint32_t C1 = 20 * w2;
const uint32_t C2 = 60 * w2;
const uint64_t xmxm = (uint64_t)stats->xm * stats->xm;
const uint64_t ymym = (uint64_t)stats->ym * stats->ym;
const int64_t xmym = (int64_t)stats->xm * stats->ym;
const int64_t sxy = (int64_t)stats->xym * N - xmym; // can be negative
const uint64_t sxx = (uint64_t)stats->xxm * N - xmxm;
const uint64_t syy = (uint64_t)stats->yym * N - ymym;
// we descale by 8 to prevent overflow during the fnum/fden multiply.
const uint64_t num_S = (2 * (uint64_t)(sxy < 0 ? 0 : sxy) + C2) >> 8;
const uint64_t den_S = (sxx + syy + C2) >> 8;
const uint64_t fnum = (2 * xmym + C1) * num_S;
const uint64_t fden = (xmxm + ymym + C1) * den_S;
const double r = (double)fnum / fden;
assert(r >= 0. && r <= 1.0);
return r;
}
double VP8SSIMFromStats(const VP8DistoStats* const stats) {
const uint32_t w = (2 * VP8_SSIM_KERNEL + 1) * (2 * VP8_SSIM_KERNEL + 1);
return SSIMCalculation(stats, w);
return SSIMCalculation(stats, kWeightSum);
}
double VP8SSIMFromStatsClipped(const VP8DistoStats* const stats) {
@ -739,16 +743,18 @@ static double SSIMGetClipped_C(const uint8_t* src1, int stride1,
src2 += ymin * stride2;
for (y = ymin; y <= ymax; ++y, src1 += stride1, src2 += stride2) {
for (x = xmin; x <= xmax; ++x) {
const int s1 = src1[x];
const int s2 = src2[x];
stats.xm += s1;
stats.ym += s2;
stats.xxm += s1 * s1;
stats.xym += s1 * s2;
stats.yym += s2 * s2;
const uint32_t w = kWeight[VP8_SSIM_KERNEL + x - xo]
* kWeight[VP8_SSIM_KERNEL + y - yo];
const uint32_t s1 = src1[x];
const uint32_t s2 = src2[x];
stats.w += w;
stats.xm += w * s1;
stats.ym += w * s2;
stats.xxm += w * s1 * s1;
stats.xym += w * s1 * s2;
stats.yym += w * s2 * s2;
}
}
stats.w = (ymax - ymin + 1) * (xmax - xmin + 1);
return VP8SSIMFromStatsClipped(&stats);
}
@ -758,13 +764,14 @@ static double SSIMGet_C(const uint8_t* src1, int stride1,
int x, y;
for (y = 0; y <= 2 * VP8_SSIM_KERNEL; ++y, src1 += stride1, src2 += stride2) {
for (x = 0; x <= 2 * VP8_SSIM_KERNEL; ++x) {
const int s1 = src1[x];
const int s2 = src2[x];
stats.xm += s1;
stats.ym += s2;
stats.xxm += s1 * s1;
stats.xym += s1 * s2;
stats.yym += s2 * s2;
const uint32_t w = kWeight[x] * kWeight[y];
const uint32_t s1 = src1[x];
const uint32_t s2 = src2[x];
stats.xm += w * s1;
stats.ym += w * s2;
stats.xxm += w * s1 * s1;
stats.xym += w * s1 * s2;
stats.yym += w * s2 * s2;
}
}
return VP8SSIMFromStats(&stats);